Survey on Deep Learning-Based Marine Object Detection

Table 2

Application of improved YOLO backbone network in maritime object detection.

Algorithms (backbone)	Datasets	Scenarios	Improved method	Effect
YOLOv2 [71]	Small ship dataset	Small ship detection	Density-based spatial clustering (DBSCAN)	AUC: 0.960; TPR: 98.3%; FPR: 3.5%
YOLOv3 [77]	SeaShip dataset	Ship detection	Loss function (GIOU)	mAP (SeaShip): 98.37%
	Buoy dataset		PANet replaces FPN	mAP (Buoy dataset): 90.58%
YOLOv2 and CNN [12]	Pascal VOC	Ship detection	—	Recall: 77.12%
	SMD			IoU: 66.69%
YOLOv3 [75]	Shanghai port surveillance video	Ship detection	—	Average acc.: 0.84
YOLOv3 [79]	SeaShip dataset	Ship detection	CBAM	mAP increase 9.6%
YOLO [123]	Self-collected	Ship detection	—	Average acc.: 92.85%
YOLOv3 tiny [124]	From Internet	Ship detection	Dense connection spatial separate conv.	LSDM average acc.: 94%; LSDM tiny: 93.5%
YOLOv3 [132]	LWIR	Object detection	—	[email protected] IoU: 0.97
				[email protected] IoU: 0.90
				[email protected] IoU: 0.29
YOLOv3 [125]	SMD; PETS 2016	Ship tracking	—	mAP: 41.2%
YOLOv2 [130]	Pascal VOC	Object detection	Pass through layer	Recall: 73.86%
	SMD	Ship detection	Transfer learning	IoU: 60.79%